Hermes Agent OS GitHub: Install Herald OS, Then Pick A Brain With Real Benchmark Scores

The Hermes Agent OS GitHub repo is github.com/iamlukethedev/Herald-OS, and it holds Herald OS, a free, open-source operating system that uses Hermes Agent as its interface.

Herald OS was built by Luke The Dev as an independent project, and it is not affiliated with Nous Research, the team that publishes Hermes Agent.

The code is released under the MIT licence, so you can download it, read every line and run it without paying for the software.

The problem it solves is simple, because instead of driving Hermes from a terminal, you get a clean desktop where you talk or type to one agent that works across your whole machine.

Every download lives on the repo's releases page, and the README walks through a route for a Mac, a spare PC, Arch, Omarchy, other Linux and a virtual machine.

In this post I will show you what is in the repo, how it works, how to install it, what every screen does and where it still falls short.

What is in the Hermes Agent OS GitHub repo?

Herald OS describes itself as an agent-native operating system, with Hermes Agent as the interface.

You talk or type to Hermes, and it works across the whole machine, asking before it changes anything.

These are the facts I checked in the repo and the GitHub API on 8 October 2026.

One naming note is worth a line here.

Nous Research codenamed Hermes Agent v0.20 the Herald Release, but Herald OS is a separate project that simply runs on Hermes.

How Herald OS works on top of Hermes

Hermes Agent is the part that thinks and acts, and Herald OS is the desktop you use to talk to it.

Herald OS does not fork Hermes, so your existing Hermes setup carries straight over.

On its first start, it links the herald-os-bridge plugin into your Hermes plugins folder and switches its tools on.

That plugin gives Hermes system tools for system info, processes, disk usage, file search, opening apps, and moving and trashing files.

Each tool has a permission tier, so reading and opening things runs straight away, changing files asks first, and trashing files or stopping a program asks every time.

Some locations, such as your SSH keys, your keychains and Hermes's own credentials, are always refused, and Settings shows the audit log of what Hermes did.

Because Herald OS sits on your one Hermes install, my automations showed up in Herald OS, Hermes Desktop and my Agent OS at the same time.

How to install Herald OS from GitHub

Every route starts with Hermes Agent and a model provider.

curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
source ~/.zshrc
hermes setup

The hermes setup step signs you in to a provider, and I cover which brain to pick further down this post.

Then pick the route that matches your machine.

I took the easy road and asked Claude to set it up for me, and it was done in about two minutes.

To change the brain later, run hermes model, or open Settings inside Herald OS.

To update on Linux, run herald-os update, and herald-os rollback takes you back.

A tour of what Herald OS can do

The dashboard says "Good morning, Julian. Your day already in motion", with voice input, today's schedule and what needs attention.

"Pick up where you left off" suggests where to continue from your recent documents, projects, chats and calendar events, and you can toggle it.

Recent screenshots from your desktop appear on the dashboard too.

The chat is normal Hermes chat with Preview and Sources, plus a new conversation button.

Spaces split work into Ideas, Work and Personal, and you can add a space per client and set a default.

Missions let Hermes plan a goal step by step, such as building out a new website.

Memory shows everything Hermes remembers, including my Telegram details, so be careful when screen-sharing.

Files covers recent, favourites, projects, downloads, shared and deleted.

Automations synced across Herald OS, Hermes Desktop and my Agent OS, including a 6am to-do recap, Japanese lessons and inbox triage.

Settings cover runtimes including Docker, background agents, remember preferences, auto-approve, activity history, fullscreen, default space and folder, do not disturb, stay awake, screen saver, dictation, themes, software access for Claude Code, Codex and LM Studio, plugins from a GitHub link, a Hermes Agent section, usage, voice engines and speeds, privacy, notifications, networks, storage and accessibility.

The Usage page in Settings shows tokens used, your plan and your token balance, and the README says it warns you when you reach 90% of your model plan.

My free Hermes Agent OS guide covers running the whole thing without paying for software.

Limits to know before you install

Herald OS is alpha software, so keep backups of anything you let an agent touch and keep approvals on.

The sidebar resizes, but the top of the interface was cut off slightly on my screen.

Connections did not sync for me, which is fair for an alpha, but it means you should not rely on them yet.

The Mac build is not notarized, so only download it from the official releases page.

Reporting issues on the Herald OS repo

Bugs go in the repo's Issues tab, with what happened, steps, platform, versions and logs.

Problems with models and providers belong on the Nous Research hermes-agent repo instead, because that is where the model plumbing lives.

Security problems go privately through the repo's Security tab.

Which brain to connect to Herald OS

Herald OS does not come with its own model, because it uses whichever model your Hermes install is set up with.

That one choice decides how smart your new operating system feels and how much it costs you to run.

To compare the options fairly, I use GoldieBench, the leaderboard where I score models on real one-shot builds.

What GoldieBench can and cannot tell you about Herald OS

I want to be straight about this before we look at numbers.

GoldieBench gives every model the same one-shot prompt to build a game, a page, a simulation or a visual, then scores the rendered result from 0 to 10.

Herald OS asks a model to do something different, which is to act as an agent through many tool calls across your whole computer.

So a GoldieBench score measures raw building ability, not how well a model runs inside Herald OS.

Herald OS itself is not a model, so it is not on the board, and I am not going to pretend otherwise.

What the board does give you is a fair, same-prompt comparison of the brains Hermes lets you connect.

The brains Herald OS can use, matched to real GoldieBench scores

The README lists the providers Hermes supports, including Nous Portal, OpenRouter, OpenAI, Anthropic and local models.

Here is how the main models behind those providers score on the live board.

Hermes MoA deserves a special mention, because it is Hermes Agent's own Mixture of Agents mode.

It scored 8.17 across 47 tasks, which shows Hermes itself can sit near the top of the board when you give it strong brains to combine.

The Claude Opus 5 vs GPT-5.6 Sol head-to-head is worth a look if you are torn between the two biggest names.

The model I actually had plugged in

In my Herald OS settings, the preferred model was GPT-6 Astra, which I had connected through a CLI.

GPT-6 Astra is not on GoldieBench yet, so I cannot give you a score for it.

That is the honest answer, and I would rather leave a gap than invent a number.

If you want a benched choice from the same family, GPT-5.6 Sol averages 8.16 across all 50 tasks.

The best brain for each part of Herald OS

Herald OS has several jobs inside it, and they do not all need the same brain.

Hermes lets you set up profiles, so you can match the brain to the job.

Automations matter most for your bill, because they run every day whether you are watching or not.

Free and local brains for Herald OS

Herald OS itself is free, and you can keep the model free as well.

Settings has a software access section where you can add local models such as LM Studio, and Hermes supports local providers.

Here is what the local models board says about the free options I have benched.

There is also LongCat-2.0, which scored 8.12, but only across 4 tasks, so the board marks it provisional.

The gap is the real story.

The best free local builder scores 7.14, the next one down scores 4.83, and the cloud leaders sit above 8.

If you want my full ranking of local brains for Hermes, read the best local model for Hermes Agent.

What each brain costs you inside Herald OS

Herald OS shows you the bill on its Usage page, which I really like.

Here is how the list prices compare against the scores.

MiniMax M3 is the value stand-out, because it sits about three tenths of a point behind Claude Opus 5 at a far lower input price.

Settings also lets you set how many background agents run, and fewer agents means a smaller bill.

My honest verdict on Herald OS and its brains

Herald OS is super clean, it synced with my existing Hermes setup quickly, and I think it looks nicer than Hermes Desktop.

It is alpha, so keep backups and keep approvals on.

If you already pay for Claude or ChatGPT, connect that first, because the board puts Claude Opus 5 at 8.27 and GPT-5.6 Sol at 8.16.

If you are choosing on value, MiniMax M3 at 7.97 is the one I would test first for daily automations.

Also On Our Network

• Agent Operatorsevery Herald OS install route with the real commands
• AI Income Desksix ways a business can use the Herald OS repo
• AI Tool Verdictmy scored review of Herald OS
• agentos.guidefitting Herald OS into an Agent OS workflow
• agentos.guidewhat the Hermes v0.20 Herald Release changed

Install Herald OS, connect the brain the board backs for each job, and you will get the most out of the Hermes Agent OS GitHub repo.